Alex Cinovoj builds AI agents and workflow automations for businesses, with a focus on getting them working reliably in production—not just in demos. Cinovoj's posts cover connecting agents to business tools, handling errors and monitoring, and choosing practical AI solutions for small and mid-sized companies.
Most useful for practical posts about AI agents and automations moving from demo to production.
Best for
AI infrastructure vendors targeting engineering leaders and small or mid-market operators.
Not for
Deck-led AI transformation programs or broad non-technical consumer lifestyle campaigns.
Key takeaways
They say the model is about 10% of a working agent.
MCP post advises one workflow, instrumenting each step, and reusing tools.
A fintech agent made 847 API calls instead of 12 and cost $6,200 in week one.
Check first: Confirm willingness and ask for formats, pricing, deliverables, and audience composition.
Main ideas
AI work is production work, and they say most people selling it cannot ship · The harness around the model is most of the work · Judge a model on messy client documents, not on leaderboards
Alex Cinovoj is known for building AI agents and workflow automations that connect to business systems and are meant to work beyond a demo.
Their LinkedIn posts explain the practical work around deployment: choosing a manageable workflow, connecting tools, handling failures, and checking costs and results.
They also write about small-business AI and emerging tools in areas such as software development and 3D modeling.
Alex Cinovoj's distinctive contribution is a builder’s focus on what happens after an AI model produces an answer.
In an MCP guide, they lay out a concrete path from a client through an MCP server and database to webhooks, then recommend starting with one workflow, instrumenting each step, and reusing tools.
In another post, they describe a logistics client whose invoices included scans, poorly lit photos, and handwriting; the remaining manual verification time helped them judge whether a newer model was worth its higher price.
They also point to limits: their post on AI implementation argues that systems fail when teams neglect integration, debugging, and production costs, while their post on hiring questions experience requirements that overlook people who have built working products.
Follow their work for practical breakdowns of AI implementation, tools, and business workflows, with examples framed around measurable operational problems rather than model capability alone.
Alex Cinovoj argues that the hard part of AI work is not building a demo but keeping an agent running on real business systems.
Read more
The key skills, they say, are scoping, integration, monitoring and cost control—not model choice.
Across their LinkedIn posts, they make the same case from several angles: the model accounts for roughly a tenth of what makes an agent work, most failures stem from process architecture, and small businesses need tools that handle messy inputs rather than the newest release.
They also discuss hiring and argue that free browser tools are displacing expensive software licences.
Their client figures and recurring claim to have shipped more than 50 production agents are their own account.
AI work is production work, and they say most people selling it cannot ship
They sort the people they see doing AI work into four failure types: consultancies billing six-figure retainers and taking six months to deliver nothing; fast prompters who can wire n8n flows but lack the debugging instincts to fix an agent at 3 a.m.; specialists who reach for custom model training when a well-prompted Claude call would do; and in-house developers who know the codebase but treat a probabilistic system like deterministic software.
Read more
The work, they say, requires a sequence: scope the workflow, build the agent, wire the integrations, instrument observability, handle edge cases and own the ROI. They call that role a production AI operator and cite “I've shipped 50+ production agents” as their standing credential.
They describe a Series B fintech that spent $120,000 over six months on an agent that worked in demos but, in production, made 847 API calls where 12 were needed.
The result: a $6,200 bill in week one and a legal escalation by week three.
They say their team rebuilt it in three weeks at about eight cents per workflow, and it is still running.
Their recurring technical claim is that the model accounts for roughly 10 percent of what makes an agent work; the rest is tools, rules, sandboxes, guardrails and observability.
Read more
They describe about 85 percent of failed AI projects as process-architecture problems, not technology problems, and frame agentic engineering as front-loading specification and harness work so each later feature costs less.
Their named stack: CrewAI crews that plan and delegate, custom Python tools exposed over MCP, n8n or Power Automate for orchestration, LangGraph with retrieval and citations, and Azure AI Foundry for governance and evaluation.
A working agent, they write, changes a record; it does not just produce a paragraph.
Three short client rollouts make the case: a 15-person Dutch firm with EUR 2M in revenue that cut admin work 60 percent and reduced response time from 24 hours to 30 seconds; a Sydney logistics firm that spent $180,000 on custom AI and cut operations headcount 30 percent after a missed customs deadline; and a Portuguese small business that invested EUR 45,000 and reports EUR 87,000 in first-year benefit.
The companies are identified by country, size and sector, not by name; the figures come from the post.
The same post credits a Google guide's co-authors, Addy Osmani, Shubham Saboo and Dr.
Judge a model on messy client documents, not on leaderboards
Their test for a model is whether it can handle a real client's bad inputs, not how it scores on a benchmark.
Read more
They state prices and weigh them against measured staff time rather than declaring a winner. Small businesses, they conclude, do not need the newest or cheapest model; they need one that works with the inputs they actually have.
A 40-person logistics client receives invoices as scanned PDFs, driver-taken photos and handwritten notes.
Automation cut the bookkeeper's 30-plus hours a week of data entry, but more than 15 hours of manual verification remained because handwriting, bad lighting and inconsistent multi-page formats defeated earlier models.
They say Claude Fable 5 handled it all, reading charts and reconstructing data from screenshots.
At $10 per million input tokens and $50 per million output—double Opus—they argue the economics work against 15 hours of bookkeeper time a week.
The embedded video is credited to Hyperagent.
The same engagement appears twice.
The first version adds benchmark detail: SWE-Bench Pro moves from 69.2% on Opus 4.8 to 80.3% on Fable 5, with a one-million-token context window.
Posted two months later, the second version drops the benchmarks but keeps the price and client outcome.
Route each task to the model it needs rather than cut AI or cut people
When a CFO frames the choice as cutting AI usage or cutting headcount, they argue that both options mean shipping less—and both are backwards.
Read more
Their alternative is automatic routing: a frontier model when the work genuinely needs one, an open-weight model such as Kimi when it does not, all within the same chat and code experience. They attribute inflated token bills to a common pattern: simple Slack summaries and quick email drafts all incur frontier-model pricing.
They describe Coworker.ai as routing each task to the appropriate model, cutting costs 82 percent—or delivering about five times more tokens for the same spend.
MCP is a shared contract for connecting agents to real tools, and simpler is often better
They define MCP as a universal way to expose tools and data to language models: an API with a shared contract that lets applications talk to each other.
Read more
Their implementation rules are concise: start with one workflow, not twelve; wrap real tools such as the database, files and CRM as MCP tools instead of writing elaborate prompts; keep glue code to 30 to 50 lines per tool; instrument the trigger, action and result; and reuse the same tools across teams rather than rebuilding them for each app. They add a limit to the recommendation: if a webhook plus SQL does the job, ship that instead.
Examples include a chain running Cursor as the MCP client into their own MCP server, then Supabase, webhooks and a Claude reply; a legacy service-desk system for a local trades business synced without vendor SDKs; and lead routing and alerts built as a single-file endpoint rather than three microservices. The guide behind the post is credited to Daily Dose of Data Science.
Free, browser-based tools are collapsing the price of expensive incumbents
Outside AI agents, they return to the same pattern: a free or open-source tool appears, they explain why it threatens a paid incumbent, then describe what their own team shipped at the same scale.
Read more
The argument targets buyers, focusing on licence costs and per-seat prices rather than technical novelty.
Pascal Editor is an open-source 3D building editor that runs in the browser on React Three Fiber and WebGPU.
It includes a building, level, wall and zone hierarchy, undo and redo, and a Next.js frontend; it has 16.9 thousand GitHub stars and an MIT licence.
They compare it with BIM seats costing $50,000 or more and say their firm shipped something similar for a client in six weeks.
They also describe GenRecon, a research release, as a Matterport replacement that turns casual smartphone video into a professional-grade mesh with no LiDAR or $500 scans.
They compare it with a property management client that had been paying about $18,000 a year across 47 short-term rentals, then switched to iPhone scans and Gaussian splatting at $1,200 a year, with guest complaints down 80 percent.
The Pascal Editor post appears twice in near-identical form. The later version adds the full colour-editing update and GitHub star count.
They argue that requiring years in a technology screens out people who can build with it, and that AI recruiting tools will not fix badly written requirements.
Read more
Such tools might at least force companies to define what “qualified” means. They say a junior who ships beats a senior who talks, and that their best hire had zero years in the stack and had simply built something that worked.
They recount a job advertisement requiring four or more years of FastAPI experience—a bar they say the framework's creator could not have met, because the project is about a year and a half old. The developer is not named.
The post appeared three times, with only punctuation and spacing differences. Two versions shift the anecdote into the third person (“They saw a job post”).
When a major platform ships a small-business product, they see an opening, not a threat to custom work: off-the-shelf workflows handle common cases while their firm builds the rest.
Read more
They argue that companies adopting now will compound their gains while those that wait play catch-up for years, and point to small businesses as a large, underserved share of the economy. Most of their higher-intent posts on this theme invite readers to message them.
They describe Anthropic's Claude for Small Business as a one-toggle install running inside QuickBooks, PayPal, HubSpot, Google Workspace and Microsoft 365, with 15 ready-to-run workflows across finance, operations, sales, marketing, HR and customer service.
They contrast it with their firm's client systems, including payroll preparation that pulls from multiple accounts and flags discrepancies, and CRM workflows that chase invoices before anyone remembers.
They cite small businesses as accounting for about 44 percent of US GDP, while many still run on spreadsheets and sticky notes.
It answers the question a newcomer actually has—what does Alex Cinovoj do, and what do they mean by "production" AI?—in a single argument.
Cinovoj divides people selling AI work into four groups that fail in different ways: large consultancies billing six-figure retainers and taking six months to deliver nothing; fast prompters who wire n8n flows but lack the debugging instincts to fix an agent that breaks at 3 AM; researchers who want to fine-tune a model when a well-prompted Claude call already works; and in-house developers who know the codebase but treat probabilistic systems like deterministic software.
They close with a client story: a Series B fintech spent about $120,000 over six months on an agent that worked in demos but, in production, made 847 API calls where 12 were needed.
The result was a $6,200 bill in week one and a legal escalation by week three.
The team rebuilt it in three weeks at roughly eight cents per workflow, and they say it is still running.
Cinovoj names the role they claim: a "Production AI Operator"—someone who scopes the workflow, builds the agent, wires integrations, instruments the system, handles edge cases, and owns the return.
For brands planning a campaign
Useful for a brief focused on reliability and cost control rather than AI enthusiasm. The specific comparison numbers can anchor a sponsored case study or webinar, but they also invite scrutiny, so plan a long-form technical follow-up rather than a short ad.
This is the clearest answer to "what would we actually be buying?"—it names the tools, the sequence, and where the work ends.
Read more
Cinovoj walks through the stack: Lovable and Cursor to get a project to about 90 percent, Claude Code for the final 10 percent, Deep Agents for open-source control, and OpenClaw or Hermes for multi-agent orchestration.
For enterprise work, they list CrewAI crews that plan and delegate, custom Python tools exposed over MCP so agents can reach internal APIs, n8n or Power Automate to connect agents to CRMs, inboxes, calendars and webhooks, LangGraph with retrieval so answers carry citations, and Microsoft's Azure AI Foundry for model governance and evaluation.
The post's central point: the output is not a paragraph but a changed record.
They list what the system is expected to handle: rate limits, validation, retries, human checkpoints, logging, and testability.
Named outcomes include intake scheduling in minutes, document triage with citations, cleaner CRM pipelines, and faster support response.
For brands planning a campaign
A natural fit for developer and agent-infrastructure tooling, where naming specific frameworks is an asset rather than jargon. A brand could sponsor the workflow-scoping step or observability layer, since Cinovoj treats instrumentation as part of the deliverable, not an add-on.
This post shows how Cinovoj argues about where AI value comes from, grounding a claim in someone else's research before adding personal client cases.
Read more
Cinovoj uses a free Google guide on agentic engineering as the spine of the argument.
Their framing: about 85 percent of failures are process architecture problems, not technology problems; the model is roughly 10 percent of what makes an agent work, while tools, rules, sandboxes, guardrails and observability make up the rest.
They cite a study of 50 small-business AI rollouts showing document processing and intake automation paying back in about 90 days, while sales forecasting and churn prediction failed at high rates for companies under 100 employees.
Three client stories follow, with numbers: a 15-person Dutch company at €2M revenue cut admin by 60 percent and moved response time from 24 hours to 30 seconds; a Sydney logistics firm spent $180,000 and cut operations headcount 30 percent, triggered by a missed customs deadline; a Portuguese small business invested €45,000 and reports €87,000 in year-one benefit.
Cinovoj credits the guide's co-authors by name.
For brands planning a campaign
A strong vehicle for a platform or research partner looking to distribute a guide, benchmark, or report to a builder audience. Cinovoj supplies the practitioner translation and small-business framing, which suits sponsors whose buyers are SMBs rather than enterprises.
To judge technical depth rather than take Cinovoj's word for it, read this post for the plumbing.
Read more
It shows what changes in a business once agents can call real tools.
Cinovoj explains MCP as a shared contract that lets apps talk to language models, then describes the chain: Cursor acting as the MCP client, the MCP server, Supabase, webhooks, and a Claude reply.
They say a legacy service-desk system for a local trades business was synced without vendor SDKs, lead routing and alerts became a single-file endpoint instead of three microservices, and the result was fewer brittle integrations.
Their five rules are short and specific: start with one workflow rather than twelve; wrap real tools such as the database, files and CRM as MCP tools instead of writing elaborate prompts; keep glue code to 30–50 lines per tool; instrument the trigger, action and result; and reuse the same tools across teams instead of rebuilding per app.
They close by saying that if a webhook and SQL do the job, ship that instead.
For brands planning a campaign
An unusually good fit for a developer-tool sponsor seeking a practical, non-hype walkthrough. The 30–50-line rule and Cursor-to-webhook chain are details that hold up to scrutiny from an engineering audience.
This shows how Cinovoj decides whether a model is worth using: not by leaderboard scores, but by whether it survives a real client's bad inputs.
Read more
Cinovoj describes a 40-person logistics client whose invoices arrive as scanned PDFs, photos from drivers, and handwritten notes.
Before automation, the bookkeeper spent over 30 hours a week on data entry; automation cut that, but 15-plus hours of manual verification remained because handwriting, poor lighting and inconsistent multi-page formats broke the models.
They report that a newer Anthropic model handled all of it, reading charts, reconstructing data from screenshots, and parsing handwriting that made other models hallucinate.
They spell out the trade-off: $10 per million input tokens and $50 per million output, which they call steep, weighed against 15 hours a week of bookkeeper time.
Their conclusion: small businesses need models that work on messy real-world inputs, not simply the newest model.
For brands planning a campaign
Alongside an earlier, nearly identical post on the same client, this shows a repeatable case-study structure an AI vendor could sponsor: real documents, a measured before-and-after, and an honest cost line. That pricing detail can build credibility with technical buyers who discount unqualified praise.
This shows how Cinovoj responds when a major platform ships a feature: they position the firm's own service as the implementation layer above it, a pattern that runs through many of their commercial posts.
Read more
Cinovoj reports that a Claude product for small business installs with one toggle and runs inside QuickBooks, PayPal, HubSpot, Google Workspace and Microsoft 365, with 15 ready-to-run workflows across finance, operations, sales, marketing, HR and customer service.
They argue this matters most for small businesses, noting they are about 44 percent of US GDP while many still run on spreadsheets.
They then name what the firm has been building—payroll prep that flags discrepancies across accounts, CRM workflows that chase invoices, and operations automations—and distinguish these from demos.
Several of their posts end with a direct-message invitation for a small or mid-market team.
For brands planning a campaign
Best suited to launch-moment campaigns where a vendor wants a builder's read on new functionality. One constraint for partners: much of Cinovoj's higher-intent content directs readers to their own services, so a sponsorship needs to fit within that implementation framing rather than compete with it.
This is the clearest example of the pattern repeated outside AI agents: a free or open-source tool appears, Cinovoj explains why it threatens a paid incumbent, and they state what the team shipped at the same scale.
Read more
Cinovoj describes Pascal Editor, an open-source 3D building editor that runs in a browser on React Three Fiber and WebGPU, with a building/level/wall/zone hierarchy, GPU-driven systems, undo and redo, and a Next.js frontend.
They note 16.9 thousand GitHub stars under an MIT license, that the latest update added full color editing for walls and furniture, and that architecture firms pay $50,000 or more per seat for comparable BIM software.
They say the firm shipped something similar for a client in six weeks and invite readers to discuss whether it could work inside their organisation.
Earlier, they made the same argument about a research lab open-sourcing a smartphone-video pipeline that competes with dedicated 3D scanning vendors, including for a property-management client where iPhone scans replaced about $18,000 a year in fees and cut guest complaints 80 percent.
For brands planning a campaign
Suited to hardware, GPU, cloud-rendering or creative-software brands looking to join an "expensive incumbent gets replaced" narrative. The open-source and cost-plus-timeline structure gives a sponsor a testable claim, though the second-hand tool details mean sponsors should expect fact-checking pressure.
Cheap, fast 3D capture against expensive incumbents3
Hiring metrics: tenure versus the ability to ship3
MCP as a practical integration pattern2
Matching models to messy real-world documents2
Token costs and routing work to the right model1
Off-the-shelf AI for small businesses1
Production agents, not demos
Across several posts, Alex Cinovoj argues that the hard part of AI work is not building a demo but keeping an agent running when it meets real systems.
Read more
They list failure types they say they see repeatedly: firms selling six-figure “AI transformation” retainers without shipping anything; fast prompters who can wire n8n flows but cannot debug an agent that breaks at 3 a.m.; specialists who reach for custom model training when a well-prompted Claude call would do; and in-house developers who treat a probabilistic system like deterministic software.
Their description of the work is a sequence: scope the workflow, build the agent, wire the integrations, instrument observability, handle edge cases and own the ROI.
They call that role a “production AI operator.”
The most detailed example is a Series B fintech that, they say, spent $120K over six months on an agent that worked in demos but made 847 API calls instead of 12 in production, producing a $6,200 bill in week one and a legal escalation by week three.
They say their team rebuilt it in three weeks at $0.08 per workflow.
Failure categories, followed by a detailed client story with dollar figures and a question inviting readers to share their worst example in the comments.
The “50+ production agents” count and the fintech rebuild are Alex Cinovoj’s own claims. Ask for third-party verification.
For brands planning a campaign
Strong fit for vendors selling observability, evaluation, agent infrastructure or integration tooling: the argument is that the model is only a small part of a working deployment. Weak fit for a sponsor message that treats AI agents as finished, plug-in products.
Alex Cinovoj’s recurring technical claim is that the model is roughly 10% of what makes an agent work; the rest is tools, rules, sandboxes, guardrails and observability.
Read more
In a post about a Google guide, they frame 85% of failed AI projects as process-architecture problems rather than technology problems, and describe agentic engineering as front-loading specification and harness work so every later feature costs less.
They back the argument with short rollouts: a 15-person Dutch firm that cut admin work 60% and moved response time from 24 hours to 30 seconds; a Sydney logistics firm that spent $180K on custom AI and cut operations headcount 30% after a missed customs deadline; and a Portuguese small business that spent about EUR 45K across inventory, support and marketing and reports EUR 87K in first-year benefits.
The posts name the stack pieces behind this work: CrewAI crews, custom Python tools exposed over MCP, n8n or Power Automate orchestration, LangGraph with retrieval and citations, and Azure AI Foundry for governance and evaluations.
Numbers from named, lightly anonymised client rollouts, paired with a stack list, turn an abstract claim about harnesses into something readers can check against their own projects.
The companies are anonymised by country, size and sector rather than named. Ask for permission to identify them or for documentation to substantiate the examples.
For brands planning a campaign
Best suited to platforms selling the non-model layer, such as orchestration, guardrails, evaluation and logging, or to a cloud provider whose governance tooling Alex Cinovoj already names. A model provider pitching model quality alone is a harder sell.
One post works through an MCP build from the inside.
Read more
Alex Cinovoj defines MCP as a universal way to expose tools and data to large language models, or as “an API with a shared contract.”
The pipeline runs Cursor as the MCP client into their own MCP server, then Supabase, webhooks and a Claude reply.
A legacy service-desk system for a local trades business was synced without vendor SDKs; lead routing and alerts became a one-file endpoint rather than three microservices.
Their implementation list: start with one workflow instead of twelve; wrap real tools (database, files, CRM) as MCP tools rather than relying on elaborate prompts; keep glue code to 30 to 50 lines per tool; instrument every trigger-action-result; and reuse the same MCP tools across teams instead of rebuilding them for each application.
They close with a note on restraint: if a webhook plus SQL does the job, ship that instead.
Definition, a short architecture diagram in text, a numbered list of implementation moves, then a limit case where MCP is the wrong choice.
For brands planning a campaign
A natural fit for integration and developer-platform brands, or a sponsor seeking a technical explainer with a concrete build sequence instead of a product pitch. The closing note on restraint also gives a sponsoring tool a credible role as an option, not a necessity.
Two posts tell the same client story: a 40-person logistics company receives invoices as scanned PDFs, photos taken by drivers and handwritten notes.
Read more
Automation cut the bookkeeper’s 30-plus hours a week of data entry, but the client still needed 15-plus hours of manual verification because handwriting, poor lighting and inconsistent multi-page formats broke the models.
Alex Cinovoj says Claude Fable 5 handled all of it, attributing that to vision capabilities that read charts, reconstruct data from screenshots and parse handwriting other models hallucinated on.
They state the price plainly: $10 per million input tokens and $50 per million output, double Opus, and argue the maths still works against 15 hours of bookkeeper time a week for a few hundred dollars of API calls.
One version adds benchmark figures: SWE-Bench Pro moving from 69.2% on Opus 4.8 to 80.3% on Fable 5, and a one-million-token context window.
Both versions end on the same point: small businesses do not need the newest or the cheapest model; they need one that works on the messy inputs they actually have.
The two posts tell the same engagement at launch and again two months later, with different benchmark details. Read them as one case, not two separate client projects.
For brands planning a campaign
Well suited to model providers and document-processing vendors willing to discuss price and benchmarks. The story format lets a sponsoring tool appear as what unblocks a real workflow, rather than as a headline claim.
A post about Coworker.ai describes a third option for teams weighing whether to cut AI usage or headcount: route each task automatically to the appropriate model, using a frontier model such as Opus when the work needs it and an open-weight model such as Kimi when it does not, within the same chat and code experience.
Read more
Alex Cinovoj cites an 82% cost reduction, or about five times more tokens for the same spend, and links inflated token bills to a common pattern: simple Slack summaries and quick email drafts all hit frontier-model pricing. They close with their usual claim of having shipped 50-plus production agents and ask readers what conversation is happening at their company.
For brands planning a campaign
Good territory for inference providers, model routers or gateways, and cloud cost-management brands. Alex Cinovoj frames AI spending as an operating decision a CFO will eventually raise, which suits a vendor presented as the cheaper third door rather than a tool to cut.
In a post on Anthropic’s Claude for Small Business, Alex Cinovoj describes a one-toggle install that puts Claude inside QuickBooks, PayPal, HubSpot, Google Workspace and Microsoft 365, with 15 ready-to-run workflows across finance, operations, sales, marketing, HR and customer service—and no custom build or six-month implementation.
Read more
They contrast that with client systems their team has been building for months, including payroll preparation that pulls from multiple accounts and flags discrepancies, CRM workflows that chase invoices before a person remembers to, and operations automations that run overnight.
They argue that the gap between companies adopting now and those waiting compounds over years.
They cite small businesses as 44% of US GDP and say most still run on spreadsheets and sticky notes, then close with a direct-message call to action.
Alex Cinovoj launches a released product, positions their own work as the custom-build version of the same outcome, then turns to an adoption argument aimed at owners rather than developers.
For brands planning a campaign
The audience is small-business owners and operators.
Brands selling packaged SMB software, finance and accounting operations, or CRM tooling fit best.
The post already carries a direct-response call to action, so a sponsored version needs one clear next step.
Cheap, fast 3D capture against expensive incumbents
Two related threads target the cost gap in 3D capture and building design.
Read more
The first covers GenRecon, an open-source research release Alex describes as a Matterport replacement: casual smartphone video in, professional-grade PBR mesh out, with no LiDAR, $500 scans or three-week turnaround.
It is built on Trellis.2, uses DINOv3 for scene understanding and overlapping chunks for room-scale consistency, and reportedly improves 16% over the prior state of the art.
Alex compares it with work their team shipped for a property management group that had been paying $18K a year in fees across 47 short-term rentals in Austin and Nashville.
Alex says iPhone scans and Gaussian splatting cut that to $1,200 a year, with same-day refresh and guest complaints down 80%.
The second thread, posted twice in near-identical form, covers Pascal Editor, a browser-based 3D building editor built with React Three Fiber and WebGPU.
It features a full building, level, wall and zone hierarchy, ECS-style architecture, Zustand undo and redo, and a Next.js frontend; it has 16.9K GitHub stars and an MIT licence.
Alex compares it with BIM seats at $50K or more and says their team shipped something similar for a client in six weeks—or under 60 days in the other version.
The two Pascal Editor posts describe the same product and stack with small edits; one adds the colour-mode update and star count. They count as one example, not two.
For brands planning a campaign
This reaches a different audience segment: architecture, real estate, construction, game development and VR buyers.
The strongest fit is 3D capture hardware or software, or a vendor willing to position itself as the fast, cheap alternative to an incumbent licence.
Claims about Alex Cinovoj’s own client builds are self-reported.
Three near-identical versions of a post argue that years of experience are a poor proxy for skill.
Read more
The anecdote centres on a job advertisement requiring four or more years of FastAPI experience—more than the person who created FastAPI could meet, since the project is only about a year and a half old.
Alex argues that AI recruiting tools will not fix badly written requirements but might force companies to define what “qualified” actually means.
Alex says their best hire had zero years in the stack and had simply built something that worked.
The post ends by asking readers for the worst job requirement they have seen.
The three posts repeat the same text with punctuation and spacing differences; they are not three separate positions.
For brands planning a campaign
Suitable for hiring, recruiting or skills-assessment brands. It also offers a lighter, non-tooling topic for sponsors looking to reach engineering leaders through a workplace theme rather than a product.
Almost every post opens with a short, blunt hook: "This robot is your AI implementation in 2026," "Someone just killed the 3D scanning industry," "Architects are going to hate this," "Your CFO just walked in with two options."
Read more
The body tends to run in one-line paragraphs, with arrow bullets for stack lists and a concrete client or product example carrying figures.
Posts about a build usually include the underlying components.
Most close with a reader question, a repost and follow request, and occasionally a direct-message keyword such as "AGENT."
A small set of signature claims recurs as shorthand for Alex Cinovoj's position, including "I've shipped 50+ production agents" and "we don't sell decks, we ship systems."
Alex frequently works from someone else's released tool or guide and credits the authors, naming Addy Osmani, Shubham Saboo, Sokratis Kartakis, Daily Dose of Data Science and Hyperagent across the posts.
Reaction posts to a release or guide alternate with stand-alone argument posts, mixing news commentary and first-person claims about client work.
Two posts contain a grammatical slip in which an anecdote shifts into the third person (“He saw a job post”); several posts also appear in multiple near-identical versions.
For brands planning a campaign
A sponsor gets a fast, specific, list-driven treatment of a product or release rather than a long-form essay. The format favours brands that can supply a concrete number, build detail or comparison to an incumbent, and puts the sponsor’s tool alongside real workflow claims.
Cinovoj writes from TechTide AI when describing client builds and inviting inbound work.
Read more
The other brands here are mentions: AI models and tools they test, endorse, or credit in their posts. The client projects they describe are not named.
TechTide AI
Disclosed partnership
Cinovoj writes as one of the people behind TechTide AI, using “we” and “our team” to describe builds and closing posts with an invitation to DM about similar projects.
Read more
They cite TechTide work such as a browser-based 3D building editor shipped in six weeks and a phone-scan replacement for a property management group's $18k-a-year 3D scanning vendor. Ask whether they own the firm, work for it, or contract for it.
Cinovoj evaluates Anthropic's Claude models against real client workflows, reporting that Claude Fable 5 read handwritten invoices, bad-lighting photos, and inconsistent multi-page formats that broke other models for a 40-employee logistics client.
Read more
They quote input and output token prices and pitch the model as the practical choice for messy business inputs.
Cinovoj promotes Coworker.ai's automatic model routing as a way to cut token spend by 82 percent, framing it as a third option for companies weighing AI budget cuts against headcount cuts.
Cinovoj repeatedly shares and credits Daily Dose of Data Science guides, describing their MCP playbook as the one they wished they had 90 days earlier and calling their agent blueprint a bible.
Alex Cinovoj writes about what it takes to keep an AI agent running after the demo ends: scoping the workflow, wiring it into real business systems, watching the token bill, and judging a model by how it handles messy client documents rather than by its leaderboard score.
Read more
Readers can start with posts that carry a full client story or a numbered build sequence, then explore shorter arguments on hiring, model pricing and free browser software.
The same themes recur in different formats, so the topics below offer a quick way to see the range.
Three creators cover similar ground from different angles.
Cole Medin writes about running autonomous coding agents as a software factory; Josue Hernandez works from vendor documentation to explain where agents break and how to gate them; Matthew Sutherland breaks down Anthropic releases and how their pricing windows change a team's plan.
Each teaches through a different kind of source than Alex Cinovoj's client-engagement stories.
Both treat AI coding agents as production systems, not demos.
Read more
Medin describes a loop where an issue goes in and a pull request comes out, with test scenarios the builder never sees and a manager agent that pings him when something is stuck.
Alex Cinovoj argues that the harness around the model, not the model itself, determines whether an agent holds up.
Both also mix writing with video walkthroughs of what they built.
Medin builds and open-sources his own autonomous software factory, teaching from the experiment while shifting toward shipping something that works out of the box. Alex Cinovoj's posts center on client engagements, with cost and ROI figures, rather than on a product of their own.
Like Alex Cinovoj, Hernandez writes about agent reliability: what an agent can reach on an account, whether an approval gate actually fails closed, and what happens when a run stops mid-task and cannot resume.
Read more
Both publish long free guides on building and gating agents, and both price routines against published rate cards.
Hernandez draws on vendor documentation and pre-release specifications, and states plainly when he had no key for a model and ran nothing live. Alex Cinovoj describes client deployments they say they shipped, with before-and-after numbers.
Sutherland writes release breakdowns for small-business and practitioner readers, covering Anthropic's Claude for Small Business workflow by workflow and mapping how a Fable 5 pricing change affects which plan makes sense.
Read more
Alex Cinovoj covers the same products for the same kind of reader. Both weigh model price against practical outcomes instead of declaring a winner.
Sutherland explains vendor releases and plan economics in structured, itemised detail, including the mechanics of inclusion windows and usage resets. Alex Cinovoj adds client deployment stories and rebuild costs, making their posts read as field reports rather than product walkthroughs.
For a campaign aimed at CTOs interested in production-grade AI systems, Alex Cinovoj offers a LinkedIn audience that already sees hands-on AI implementation content.
Read more
The account had 64,736 followers as of 8 October 2026.
Posts cover production AI agents, MCP integrations, SMB automation, model selection, document processing, 3D scanning, browser-based building tools, and hiring.
These topics align with the campaign, but the posts do not show whether followers are CTOs, AI engineers, or operations leaders.
Ask for audience composition and recent results on comparable topics, depth, and formats—specifically technical architecture breakdowns, production failure and fix examples, and tool comparisons.
Ask whether the creator has audience survey results covering job titles, seniority, and industry, or platform demographics if available.
No sponsorship or brand-partnership offering is documented in Alex Cinovoj's posts.
Read more
The commercial activity visible there is a services business: posts describe client AI and automation builds delivered through TechTide AI and invite readers to send a direct message to request a blueprint or book a call.
They do not set out sponsorship formats, deliverables, pricing, or audience data.
Brands should ask by direct message whether Cinovoj is open to a partnership and negotiate scope, terms, and pricing individually.
A campaign aimed at engineering leaders and small or mid-market operators for an AI infrastructure vendor selling observability, evaluation, guardrails, or integration tooling.
Read more
The subject matter fits Cinovoj's posts, which repeatedly argue that the model is roughly 10% of a working agent, while tools, rules, sandboxes, guardrails, and observability make up the rest.
They frame about 85% of failed AI projects as process-architecture problems.
The MCP post lays out a practical integration chain (Cursor MCP client into an MCP server, Supabase, webhooks, Claude reply) and rules such as starting with one workflow, wrapping real tools, keeping glue code to 30 to 50 lines, and instrumenting trigger-action-result.
The production-agent post gives a failure taxonomy and a client story: a Series B fintech spent about $120,000 over six months on an agent that worked in demos but made 847 API calls instead of 12, produced a $6,200 bill in week one, and escalated legally by week three; they say the rebuild took three weeks at $0.08 per task.
The token-routing post argues simple tasks should not all hit frontier-model rates and cites an 82% cost reduction.
That supports a campaign about keeping agents running after the demo, not model hype.
Ask for audience composition by job title and sponsorship formats, rates, and deliverables before agreeing to a paid campaign; the posts do not provide them.
When to choose a specialist
A campaign aimed at a consultancy or platform selling a six-figure AI transformation programme built around long timelines, slide decks, and demos rather than working systems is a poor fit.
Read more
Cinovoj's strongest recurring theme is the opposite: they describe six-figure retainers, six-month timelines, fifty-slide decks, and zero production deployments as the first failure group, contrasting that with shipping agents.
They also criticise in-house developers who treat probabilistic systems deterministically and specialists who reach for custom model training where a well-prompted call would work.
Asking them to endorse deck-led transformation would run against their public argument.
A consumer lifestyle or general-interest brand seeking a broad non-technical audience is also a poor fit: their posts focus on production AI agents, MCP, token costs, SMB automation, 3D capture, and hiring.
The audience section measures a 64,736-follower LinkedIn account but provides no demographics.
Ask for audience data before committing to a consumer campaign.
Their posts also describe no sponsorship formats; ask for a media kit or agree campaign terms and deliverables before proceeding.
A worked campaign example
Take the campaign for an AI infrastructure vendor.
Campaign angle: the 3 AM test, or what keeps a production agent standing.
Mirror Cinovoj's blunt LinkedIn openings with the falling-robot premise: an agent that works in a demo, then meets real systems.
Open with the Series B fintech numbers - $120,000 spent, six months in, 847 API calls instead of 12, $6,200 in week one, legal escalation by week three - and state that the rebuild ran at $0.08 per task.
Explain the harness around the model: tools, rules, sandboxes, guardrails, observability, validation, retries, human checkpoints, logging, and testability.
Show the MCP integration path in concrete steps: start with one task, wrap the database, files, and CRM as MCP tools, keep glue code to 30 to 50 lines per tool, instrument trigger-action-result, and reuse tools across teams.
Add the routing and cost layer: simple Slack summaries and quick email drafts should not all hit frontier-model rates; route tasks to the right model and use the claimed 82% reduction as the discussion point.
Close with a reader question and one direct next step, consistent with their usual direct-message keyword or book-a-call invitation.
The infrastructure vendor supplies the observability, evaluation, or guardrail layer the angle treats as most of the work.
Do not frame the sponsor as the reason the agent works; frame it as how teams see failures before 3 AM.
Ask Cinovoj to agree scope, terms, and deliverables through a bespoke direct-message conversation, since their posts describe no sponsorship programme.
Measure whether technical readers can describe one harness component they are missing after reading.
Choose Cole Medin over Alex Cinovoj when the campaign needs to show a team building the harness, not just explain why demos fail.
Read more
Medin spent a year running a software factory: a GitHub issue goes in, a pull request comes out.
Each lap builds one issue, tests the app as a user would, reviews the code and merges into staging only when CI is green.
Laps run in parallel, and a manager agent pings him when something is genuinely stuck.
Two choices map directly to what an infrastructure sponsor sells.
He plans each release with holdout test scenarios the builder never sees, because agents that can read the tests will engineer code just to pass them.
Anything that is off becomes a new issue the factory picks up on its own.
He is now building an open source AI software factory and says it is his top priority alongside Archon.
For an evaluation or test-infrastructure sponsor whose angle is teach-the-engineer-to-build-it, Medin speaks to the reader more directly than Cinovoj.
Medin's posts focus on building software with agents, not keeping a deployed agent running inside someone else's business systems.
Cinovoj provides the operating-room detail Medin does not: a Series B fintech at roughly $120,000 over six months, an agent making 847 API calls instead of 12, a $6,200 first-week bill, legal escalation by week three, and a rebuild at $0.08 per workflow.
Medin's case rests on two posts, from September 3 and October 16, 2026, covering one project and its stated next step rather than a range of client work.
For monitoring, cost control, or surviving a 3 AM break in a live workflow, Cinovoj's material is the better base; Medin is a supporting voice.
Choose Josue Hernandez when the campaign's honest claim is diligence: read the vendor's own documentation before trusting the tool.
Read more
On Grok Bot, he found a line on the security page—“do not use separate Bots as a security boundary”—that contradicts the setup most tutorials recommend.
Every bot on the account shares one cloud computer, one filesystem and one set of browser sessions, so signing a single bot into a CRM gives that session to every bot, including ones created later.
On the OpenAI side, he found that the Agents SDK wraps approval decisions in bool(), so a predicate returning None becomes False; False reads as no approval needed, the guarded tool fires and nobody is asked.
He filed four issues in a week.
He then built the boring part first, with a stop handler and a gate that fails closed, and says on page one that he had no key for the model and nothing in it was run live.
For an observability, guardrail or auth-layer sponsor, he delivers the campaign's argument through his visible habit of checking the primary source and disclosing what has not been tested.
His posts cover documentation forensics and giveaway guides, not client outcomes.
Cinovoj can point to a rebuild that shipped and held at $0.08 per workflow; Hernandez explains the failure mode and safe default.
A sponsor seeking proof of results in production should ask for that evidence elsewhere.
Hernandez also writes for operators working inside a named platform, not buyers weighing a category, making his audience narrower than a campaign targeting engineering leaders and small or mid-market operators.
The comparison rests on two posts, from August 26 and September 11, 2026; ask for more examples before presenting him as a general authority on production AI.
Use Matthew Sutherland when the campaign turns on the buying and cost decision rather than the operating one.
Read more
His Claude for Small Business post explains what the package is, which tools it connects to—QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace and Microsoft 365—what the 15 starter workflows cover across finance, operations, sales, marketing, HR and customer service, and where the guardrails sit: every workflow is user-initiated, existing app permissions still apply, and no training on business data happens by default on Team and Enterprise plans.
His Fable 5 post tackles the pricing choice: why 20x is the right tier for a build sprint rather than 5x, how capacity meters in rolling five-hour windows and does not roll forward, why the backlog has to be queued before the window opens, and how strong caching and heavy repo re-reads land the payoff in very different places.
If the sponsor's product is bought on procurement or cost reasoning, or the campaign needs someone who explains vendor packaging to small-business owners, Sutherland is closer to the decision than Cinovoj.
Sutherland analyses published products and rate cards rather than reporting client work, so a campaign promising demonstrated results in live systems would be on the wrong basis.
His two posts, from May 18 and June 10, 2026, predate later model releases; the pricing specifics age fast, so ask for a dated recheck before presenting them as current.
Cinovoj remains the better base for engineers who will run the agent and need answers on integration, failure modes and operating cost, rather than which plan to buy or when the included window closes.
Alex Cinovoj's 14 LinkedIn posts cover agent builds, MCP integrations, production failures, model testing against client documents, small-business AI adoption, hiring metrics and browser-based 3D tools.
The posts carry no publication dates, so the material is grouped by theme rather than presented chronologically, and repeated posts appear as one example.
The "50+ production agents" claim, client figures and portfolio counts are claims made in the posts; the posts do not name clients or include independently verified credentials, engagement data, audience demographics, sponsor disclosures or video transcripts.
